Generate Your Own Industry Dataset

• ✓Build a seeded sampler and reference tables, so the record is the label and the whole corpus regenerates from a seed
• ✓Build a renderer layer that puts a local LLM behind one interface, and measure what it costs to run on your own hardware
• ✓Build a schema-constrained extraction call with nullable fields, so a model can say a value is absent instead of inventing it
• ✓Build a verifier that tells a missing value from a contradicted one, and reject every render that contradicts its record
• ✓Build a catalogue of defects whose flags are true by construction, and report the rates the corpus actually contains
• ✓Build the measures that tell you what a corpus holds, and aim one at a named failure instead of trusting a general number
• ✓Build a pipeline, a split by customer and a corpus card, so a teammate can regenerate and trust the dataset
• ✓Build a domain-agnostic core and a second industry pack, and prove with tests that the core never imports a pack

Drill 1 of the Freight & Logistics track. Start here, because drills 2 to 5 all run on the dataset you build in this one. You build a small, tested Python package that produces realistic, labelled data for an AI system when real customer data cannot be used, which is where almost every enterprise AI project begins.